Spheron

Spheron gives Claude its own GPU machine. Tell it the experiment, and it picks a machine, sets it up, runs the work and reports back. Claude compares GPU offers from several clouds and picks the cheapest one that fits the job, from a single RTX 4090 to 8x H100, or a large CPU machine when that suits better. You see the price before anything launches. What people use it for: - Run evals. "Run this eval suite on an H100 and give me the scores." Claude launches the machine, runs it and brings back the results. - Sweep settings overnight. "Try six quantization configs on an A100 and compare speed and accuracy." Come back in the morning to a table. - Train or fine-tune. Claude pulls your dataset and code, starts the run in the background, and checks on it whenever you ask. - Benchmark across hardware. Run the same test on an A6000, an RTX PRO 6000 and an H100, then pick the best value. - Process big batches. Transcribe audio, render images or crunch a dataset on a GPU instead of your laptop. - Try a new open model. "Set up Qwen on a GPU so I can test it." Claude installs everything and runs it. How it works: Claude launches the machine and then works on it directly. It runs shell commands, reads and writes files, and keeps jobs of up to 24 hours running in the background while you do something else. When a run finishes, it can stop the machine so you stop paying. You never need SSH, but you can log in yourself at any time. You stay in control. Claude shows the price and asks before anything that costs money, deletes data or grants access, and only goes ahead once you say yes. You set an hourly spend limit for the connection on spheron.ai, and Claude can't raise it. A stopped machine keeps its files and costs only a little for storage. Good to know: machines are billed by the hour at the price Claude shows before launch. The connector sees only your own Spheron account, and you sign in with OAuth.